Cheyenne Super Computer Auction

Original Article ↗ Hacker News Discussion ↗

A decommissioned 2016-era Cheyenne supercomputer, once ranked 20th on the Top500 list, was auctioned by the U.S. government and ultimately sold for about $480,000, roughly the value of its CPUs alone. Commenters examine whether it makes economic sense to buy such a 5.34-petaflop, 1.7MW system for reuse or resale, weighing the scrap value of thousands of Xeon processors and hundreds of terabytes of DDR4 RAM against the high costs of transport, power, and decommissioning. They also contrast its architecture and liquid-cooled HPC design with modern GPU-based clusters, noting that contemporary hardware can often match or exceed its performance at a fraction of the energy and operating expense.

System and Auction Overview

  • Cheyenne is a 2016 SGI ICE XA supercomputer: 4,032 dual-socket nodes (8,064 E5‑2697v4 CPUs, 145,152 cores), 313 TB DDR4, EDR InfiniBand hypercube topology, peak ~5.34 PFLOPS FP64, ~1.7 MW power draw.
  • Racks are heavy (up to ~47 tons total system weight); specially cooled E‑cells and water-cooled racks.
  • Reserve price reported as $100k; final sale reported around $480k, roughly the then-estimated CPU resale value.

Part-Out Economics and Practical Constraints

  • Many see the real value in CPUs and RAM: per-CPU eBay prices imply hundreds of thousands of dollars in potential resale.
  • Counterpoints:
    • Flooding the market could depress prices, and demand for 2016-era Xeons is limited and slow-moving.
    • Significant costs for professional movers, insurance, trucking, warehousing, and labor to disassemble, test, and list parts; some estimates put moving plus labor near six figures.
    • Much of the non-CPU gear (cooling units, custom chassis) may be near scrap value.

Possible Buyers and Uses

  • Likely buyers: recyclers/part-out resellers, low-budget academic groups, niche hosting or homelab enthusiasts.
  • Reassembling the full cluster is seen as unlikely due to power, cooling, and licensing/management complexity.
  • Some wish it could go to a museum, but maintaining and powering it even occasionally is a major burden.

Efficiency, Obsolescence, and Comparisons

  • Several argue it’s uneconomical to run: 1.7 MW implies thousands of dollars per day in electricity, while modern CPUs/GPUs or MI300X-class accelerators can match or exceed FP64 throughput at a fraction of power.
  • Others note its value as a general-purpose x86 cluster with mature MPI/Linux tooling, versus the programming effort to exploit modern GPUs.

Cooling Design and Reliability

  • Liquid-assisted cooling is framed as standard in large HPC for density and efficiency versus massive air systems.
  • Decommissioning was partly driven by failing quick-disconnects causing water spray and aging RAM, reinforcing the maintenance burden.

Nostalgia and Supercomputing Culture

  • Multiple comments express fascination and nostalgia for big iron (SGI, Cray, Origin/NUMA systems) and regret over the difficulty of preserving such machines.
  • There’s tension between viewing it as a cool artifact and acknowledging it as outdated, power-hungry technology best treated as parts.